A connection puzzle that listens to your answer

Many connection games ask you to select from a fixed set of choices. Connecdle lets you describe the relationship yourself. That small change makes the puzzle more expressive: recognising an idea is only the beginning, and putting it into words becomes part of the solve.

There is no single magic phrase. Clear paraphrases can succeed, while answers that identify only a broad category can receive close feedback and another chance.

The goal is a fair leap of lateral thinking rather than specialist trivia. A satisfying puzzle should use familiar clues, have a defensible intended connection, and feel obvious in hindsight. Specificity matters because it separates an accidental category from the reason those exact four clues were chosen.

Connecdle uses a compact language model to compare a player’s guess with the puzzle’s intended ideas. The model is hosted only for scoring; it does not write the daily answer during play. Guess outcomes can be reviewed to find false accepts, false rejects, and natural paraphrases that deserve better treatment in future model versions.

Automated judgement will never be perfect. The aim is to make it useful, transparent through correct/close/incorrect feedback, and steadily more reliable as real language replaces synthetic training examples.

One new puzzle appears at 00:00 UTC. There are no accounts, streak pressure, or endless levels. You get five thoughtful attempts, a spoiler-free result to share, and then a fresh connection the next day.